2 * Copyright (C) 2012 - Virtual Open Systems and Columbia University
3 * Author: Christoffer Dall <c.dall@virtualopensystems.com>
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License, version 2, as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
19 #include <linux/mman.h>
20 #include <linux/kvm_host.h>
22 #include <linux/hugetlb.h>
23 #include <linux/sched/signal.h>
24 #include <trace/events/kvm.h>
25 #include <asm/pgalloc.h>
26 #include <asm/cacheflush.h>
27 #include <asm/kvm_arm.h>
28 #include <asm/kvm_mmu.h>
29 #include <asm/kvm_mmio.h>
30 #include <asm/kvm_ras.h>
31 #include <asm/kvm_asm.h>
32 #include <asm/kvm_emulate.h>
37 static pgd_t
*boot_hyp_pgd
;
38 static pgd_t
*hyp_pgd
;
39 static pgd_t
*merged_hyp_pgd
;
40 static DEFINE_MUTEX(kvm_hyp_pgd_mutex
);
42 static unsigned long hyp_idmap_start
;
43 static unsigned long hyp_idmap_end
;
44 static phys_addr_t hyp_idmap_vector
;
46 static unsigned long io_map_base
;
48 #define hyp_pgd_order get_order(PTRS_PER_PGD * sizeof(pgd_t))
50 #define KVM_S2PTE_FLAG_IS_IOMAP (1UL << 0)
51 #define KVM_S2_FLAG_LOGGING_ACTIVE (1UL << 1)
53 static bool memslot_is_logging(struct kvm_memory_slot
*memslot
)
55 return memslot
->dirty_bitmap
&& !(memslot
->flags
& KVM_MEM_READONLY
);
59 * kvm_flush_remote_tlbs() - flush all VM TLB entries for v7/8
60 * @kvm: pointer to kvm structure.
62 * Interface to HYP function to flush all VM TLB entries
64 void kvm_flush_remote_tlbs(struct kvm
*kvm
)
66 kvm_call_hyp(__kvm_tlb_flush_vmid
, kvm
);
69 static void kvm_tlb_flush_vmid_ipa(struct kvm
*kvm
, phys_addr_t ipa
)
71 kvm_call_hyp(__kvm_tlb_flush_vmid_ipa
, kvm
, ipa
);
75 * D-Cache management functions. They take the page table entries by
76 * value, as they are flushing the cache using the kernel mapping (or
79 static void kvm_flush_dcache_pte(pte_t pte
)
81 __kvm_flush_dcache_pte(pte
);
84 static void kvm_flush_dcache_pmd(pmd_t pmd
)
86 __kvm_flush_dcache_pmd(pmd
);
89 static void kvm_flush_dcache_pud(pud_t pud
)
91 __kvm_flush_dcache_pud(pud
);
94 static bool kvm_is_device_pfn(unsigned long pfn
)
96 return !pfn_valid(pfn
);
100 * stage2_dissolve_pmd() - clear and flush huge PMD entry
101 * @kvm: pointer to kvm structure.
103 * @pmd: pmd pointer for IPA
105 * Function clears a PMD entry, flushes addr 1st and 2nd stage TLBs.
107 static void stage2_dissolve_pmd(struct kvm
*kvm
, phys_addr_t addr
, pmd_t
*pmd
)
109 if (!pmd_thp_or_huge(*pmd
))
113 kvm_tlb_flush_vmid_ipa(kvm
, addr
);
114 put_page(virt_to_page(pmd
));
118 * stage2_dissolve_pud() - clear and flush huge PUD entry
119 * @kvm: pointer to kvm structure.
121 * @pud: pud pointer for IPA
123 * Function clears a PUD entry, flushes addr 1st and 2nd stage TLBs.
125 static void stage2_dissolve_pud(struct kvm
*kvm
, phys_addr_t addr
, pud_t
*pudp
)
127 if (!stage2_pud_huge(kvm
, *pudp
))
130 stage2_pud_clear(kvm
, pudp
);
131 kvm_tlb_flush_vmid_ipa(kvm
, addr
);
132 put_page(virt_to_page(pudp
));
135 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache
*cache
,
140 BUG_ON(max
> KVM_NR_MEM_OBJS
);
141 if (cache
->nobjs
>= min
)
143 while (cache
->nobjs
< max
) {
144 page
= (void *)__get_free_page(PGALLOC_GFP
);
147 cache
->objects
[cache
->nobjs
++] = page
;
152 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache
*mc
)
155 free_page((unsigned long)mc
->objects
[--mc
->nobjs
]);
158 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache
*mc
)
162 BUG_ON(!mc
|| !mc
->nobjs
);
163 p
= mc
->objects
[--mc
->nobjs
];
167 static void clear_stage2_pgd_entry(struct kvm
*kvm
, pgd_t
*pgd
, phys_addr_t addr
)
169 pud_t
*pud_table __maybe_unused
= stage2_pud_offset(kvm
, pgd
, 0UL);
170 stage2_pgd_clear(kvm
, pgd
);
171 kvm_tlb_flush_vmid_ipa(kvm
, addr
);
172 stage2_pud_free(kvm
, pud_table
);
173 put_page(virt_to_page(pgd
));
176 static void clear_stage2_pud_entry(struct kvm
*kvm
, pud_t
*pud
, phys_addr_t addr
)
178 pmd_t
*pmd_table __maybe_unused
= stage2_pmd_offset(kvm
, pud
, 0);
179 VM_BUG_ON(stage2_pud_huge(kvm
, *pud
));
180 stage2_pud_clear(kvm
, pud
);
181 kvm_tlb_flush_vmid_ipa(kvm
, addr
);
182 stage2_pmd_free(kvm
, pmd_table
);
183 put_page(virt_to_page(pud
));
186 static void clear_stage2_pmd_entry(struct kvm
*kvm
, pmd_t
*pmd
, phys_addr_t addr
)
188 pte_t
*pte_table
= pte_offset_kernel(pmd
, 0);
189 VM_BUG_ON(pmd_thp_or_huge(*pmd
));
191 kvm_tlb_flush_vmid_ipa(kvm
, addr
);
192 free_page((unsigned long)pte_table
);
193 put_page(virt_to_page(pmd
));
196 static inline void kvm_set_pte(pte_t
*ptep
, pte_t new_pte
)
198 WRITE_ONCE(*ptep
, new_pte
);
202 static inline void kvm_set_pmd(pmd_t
*pmdp
, pmd_t new_pmd
)
204 WRITE_ONCE(*pmdp
, new_pmd
);
208 static inline void kvm_pmd_populate(pmd_t
*pmdp
, pte_t
*ptep
)
210 kvm_set_pmd(pmdp
, kvm_mk_pmd(ptep
));
213 static inline void kvm_pud_populate(pud_t
*pudp
, pmd_t
*pmdp
)
215 WRITE_ONCE(*pudp
, kvm_mk_pud(pmdp
));
219 static inline void kvm_pgd_populate(pgd_t
*pgdp
, pud_t
*pudp
)
221 WRITE_ONCE(*pgdp
, kvm_mk_pgd(pudp
));
226 * Unmapping vs dcache management:
228 * If a guest maps certain memory pages as uncached, all writes will
229 * bypass the data cache and go directly to RAM. However, the CPUs
230 * can still speculate reads (not writes) and fill cache lines with
233 * Those cache lines will be *clean* cache lines though, so a
234 * clean+invalidate operation is equivalent to an invalidate
235 * operation, because no cache lines are marked dirty.
237 * Those clean cache lines could be filled prior to an uncached write
238 * by the guest, and the cache coherent IO subsystem would therefore
239 * end up writing old data to disk.
241 * This is why right after unmapping a page/section and invalidating
242 * the corresponding TLBs, we call kvm_flush_dcache_p*() to make sure
243 * the IO subsystem will never hit in the cache.
245 * This is all avoided on systems that have ARM64_HAS_STAGE2_FWB, as
246 * we then fully enforce cacheability of RAM, no matter what the guest
249 static void unmap_stage2_ptes(struct kvm
*kvm
, pmd_t
*pmd
,
250 phys_addr_t addr
, phys_addr_t end
)
252 phys_addr_t start_addr
= addr
;
253 pte_t
*pte
, *start_pte
;
255 start_pte
= pte
= pte_offset_kernel(pmd
, addr
);
257 if (!pte_none(*pte
)) {
258 pte_t old_pte
= *pte
;
260 kvm_set_pte(pte
, __pte(0));
261 kvm_tlb_flush_vmid_ipa(kvm
, addr
);
263 /* No need to invalidate the cache for device mappings */
264 if (!kvm_is_device_pfn(pte_pfn(old_pte
)))
265 kvm_flush_dcache_pte(old_pte
);
267 put_page(virt_to_page(pte
));
269 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
271 if (stage2_pte_table_empty(kvm
, start_pte
))
272 clear_stage2_pmd_entry(kvm
, pmd
, start_addr
);
275 static void unmap_stage2_pmds(struct kvm
*kvm
, pud_t
*pud
,
276 phys_addr_t addr
, phys_addr_t end
)
278 phys_addr_t next
, start_addr
= addr
;
279 pmd_t
*pmd
, *start_pmd
;
281 start_pmd
= pmd
= stage2_pmd_offset(kvm
, pud
, addr
);
283 next
= stage2_pmd_addr_end(kvm
, addr
, end
);
284 if (!pmd_none(*pmd
)) {
285 if (pmd_thp_or_huge(*pmd
)) {
286 pmd_t old_pmd
= *pmd
;
289 kvm_tlb_flush_vmid_ipa(kvm
, addr
);
291 kvm_flush_dcache_pmd(old_pmd
);
293 put_page(virt_to_page(pmd
));
295 unmap_stage2_ptes(kvm
, pmd
, addr
, next
);
298 } while (pmd
++, addr
= next
, addr
!= end
);
300 if (stage2_pmd_table_empty(kvm
, start_pmd
))
301 clear_stage2_pud_entry(kvm
, pud
, start_addr
);
304 static void unmap_stage2_puds(struct kvm
*kvm
, pgd_t
*pgd
,
305 phys_addr_t addr
, phys_addr_t end
)
307 phys_addr_t next
, start_addr
= addr
;
308 pud_t
*pud
, *start_pud
;
310 start_pud
= pud
= stage2_pud_offset(kvm
, pgd
, addr
);
312 next
= stage2_pud_addr_end(kvm
, addr
, end
);
313 if (!stage2_pud_none(kvm
, *pud
)) {
314 if (stage2_pud_huge(kvm
, *pud
)) {
315 pud_t old_pud
= *pud
;
317 stage2_pud_clear(kvm
, pud
);
318 kvm_tlb_flush_vmid_ipa(kvm
, addr
);
319 kvm_flush_dcache_pud(old_pud
);
320 put_page(virt_to_page(pud
));
322 unmap_stage2_pmds(kvm
, pud
, addr
, next
);
325 } while (pud
++, addr
= next
, addr
!= end
);
327 if (stage2_pud_table_empty(kvm
, start_pud
))
328 clear_stage2_pgd_entry(kvm
, pgd
, start_addr
);
332 * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
333 * @kvm: The VM pointer
334 * @start: The intermediate physical base address of the range to unmap
335 * @size: The size of the area to unmap
337 * Clear a range of stage-2 mappings, lowering the various ref-counts. Must
338 * be called while holding mmu_lock (unless for freeing the stage2 pgd before
339 * destroying the VM), otherwise another faulting VCPU may come in and mess
340 * with things behind our backs.
342 static void unmap_stage2_range(struct kvm
*kvm
, phys_addr_t start
, u64 size
)
345 phys_addr_t addr
= start
, end
= start
+ size
;
348 assert_spin_locked(&kvm
->mmu_lock
);
349 WARN_ON(size
& ~PAGE_MASK
);
351 pgd
= kvm
->arch
.pgd
+ stage2_pgd_index(kvm
, addr
);
354 * Make sure the page table is still active, as another thread
355 * could have possibly freed the page table, while we released
358 if (!READ_ONCE(kvm
->arch
.pgd
))
360 next
= stage2_pgd_addr_end(kvm
, addr
, end
);
361 if (!stage2_pgd_none(kvm
, *pgd
))
362 unmap_stage2_puds(kvm
, pgd
, addr
, next
);
364 * If the range is too large, release the kvm->mmu_lock
365 * to prevent starvation and lockup detector warnings.
368 cond_resched_lock(&kvm
->mmu_lock
);
369 } while (pgd
++, addr
= next
, addr
!= end
);
372 static void stage2_flush_ptes(struct kvm
*kvm
, pmd_t
*pmd
,
373 phys_addr_t addr
, phys_addr_t end
)
377 pte
= pte_offset_kernel(pmd
, addr
);
379 if (!pte_none(*pte
) && !kvm_is_device_pfn(pte_pfn(*pte
)))
380 kvm_flush_dcache_pte(*pte
);
381 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
384 static void stage2_flush_pmds(struct kvm
*kvm
, pud_t
*pud
,
385 phys_addr_t addr
, phys_addr_t end
)
390 pmd
= stage2_pmd_offset(kvm
, pud
, addr
);
392 next
= stage2_pmd_addr_end(kvm
, addr
, end
);
393 if (!pmd_none(*pmd
)) {
394 if (pmd_thp_or_huge(*pmd
))
395 kvm_flush_dcache_pmd(*pmd
);
397 stage2_flush_ptes(kvm
, pmd
, addr
, next
);
399 } while (pmd
++, addr
= next
, addr
!= end
);
402 static void stage2_flush_puds(struct kvm
*kvm
, pgd_t
*pgd
,
403 phys_addr_t addr
, phys_addr_t end
)
408 pud
= stage2_pud_offset(kvm
, pgd
, addr
);
410 next
= stage2_pud_addr_end(kvm
, addr
, end
);
411 if (!stage2_pud_none(kvm
, *pud
)) {
412 if (stage2_pud_huge(kvm
, *pud
))
413 kvm_flush_dcache_pud(*pud
);
415 stage2_flush_pmds(kvm
, pud
, addr
, next
);
417 } while (pud
++, addr
= next
, addr
!= end
);
420 static void stage2_flush_memslot(struct kvm
*kvm
,
421 struct kvm_memory_slot
*memslot
)
423 phys_addr_t addr
= memslot
->base_gfn
<< PAGE_SHIFT
;
424 phys_addr_t end
= addr
+ PAGE_SIZE
* memslot
->npages
;
428 pgd
= kvm
->arch
.pgd
+ stage2_pgd_index(kvm
, addr
);
430 next
= stage2_pgd_addr_end(kvm
, addr
, end
);
431 if (!stage2_pgd_none(kvm
, *pgd
))
432 stage2_flush_puds(kvm
, pgd
, addr
, next
);
433 } while (pgd
++, addr
= next
, addr
!= end
);
437 * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
438 * @kvm: The struct kvm pointer
440 * Go through the stage 2 page tables and invalidate any cache lines
441 * backing memory already mapped to the VM.
443 static void stage2_flush_vm(struct kvm
*kvm
)
445 struct kvm_memslots
*slots
;
446 struct kvm_memory_slot
*memslot
;
449 idx
= srcu_read_lock(&kvm
->srcu
);
450 spin_lock(&kvm
->mmu_lock
);
452 slots
= kvm_memslots(kvm
);
453 kvm_for_each_memslot(memslot
, slots
)
454 stage2_flush_memslot(kvm
, memslot
);
456 spin_unlock(&kvm
->mmu_lock
);
457 srcu_read_unlock(&kvm
->srcu
, idx
);
460 static void clear_hyp_pgd_entry(pgd_t
*pgd
)
462 pud_t
*pud_table __maybe_unused
= pud_offset(pgd
, 0UL);
464 pud_free(NULL
, pud_table
);
465 put_page(virt_to_page(pgd
));
468 static void clear_hyp_pud_entry(pud_t
*pud
)
470 pmd_t
*pmd_table __maybe_unused
= pmd_offset(pud
, 0);
471 VM_BUG_ON(pud_huge(*pud
));
473 pmd_free(NULL
, pmd_table
);
474 put_page(virt_to_page(pud
));
477 static void clear_hyp_pmd_entry(pmd_t
*pmd
)
479 pte_t
*pte_table
= pte_offset_kernel(pmd
, 0);
480 VM_BUG_ON(pmd_thp_or_huge(*pmd
));
482 pte_free_kernel(NULL
, pte_table
);
483 put_page(virt_to_page(pmd
));
486 static void unmap_hyp_ptes(pmd_t
*pmd
, phys_addr_t addr
, phys_addr_t end
)
488 pte_t
*pte
, *start_pte
;
490 start_pte
= pte
= pte_offset_kernel(pmd
, addr
);
492 if (!pte_none(*pte
)) {
493 kvm_set_pte(pte
, __pte(0));
494 put_page(virt_to_page(pte
));
496 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
498 if (hyp_pte_table_empty(start_pte
))
499 clear_hyp_pmd_entry(pmd
);
502 static void unmap_hyp_pmds(pud_t
*pud
, phys_addr_t addr
, phys_addr_t end
)
505 pmd_t
*pmd
, *start_pmd
;
507 start_pmd
= pmd
= pmd_offset(pud
, addr
);
509 next
= pmd_addr_end(addr
, end
);
510 /* Hyp doesn't use huge pmds */
512 unmap_hyp_ptes(pmd
, addr
, next
);
513 } while (pmd
++, addr
= next
, addr
!= end
);
515 if (hyp_pmd_table_empty(start_pmd
))
516 clear_hyp_pud_entry(pud
);
519 static void unmap_hyp_puds(pgd_t
*pgd
, phys_addr_t addr
, phys_addr_t end
)
522 pud_t
*pud
, *start_pud
;
524 start_pud
= pud
= pud_offset(pgd
, addr
);
526 next
= pud_addr_end(addr
, end
);
527 /* Hyp doesn't use huge puds */
529 unmap_hyp_pmds(pud
, addr
, next
);
530 } while (pud
++, addr
= next
, addr
!= end
);
532 if (hyp_pud_table_empty(start_pud
))
533 clear_hyp_pgd_entry(pgd
);
536 static unsigned int kvm_pgd_index(unsigned long addr
, unsigned int ptrs_per_pgd
)
538 return (addr
>> PGDIR_SHIFT
) & (ptrs_per_pgd
- 1);
541 static void __unmap_hyp_range(pgd_t
*pgdp
, unsigned long ptrs_per_pgd
,
542 phys_addr_t start
, u64 size
)
545 phys_addr_t addr
= start
, end
= start
+ size
;
549 * We don't unmap anything from HYP, except at the hyp tear down.
550 * Hence, we don't have to invalidate the TLBs here.
552 pgd
= pgdp
+ kvm_pgd_index(addr
, ptrs_per_pgd
);
554 next
= pgd_addr_end(addr
, end
);
556 unmap_hyp_puds(pgd
, addr
, next
);
557 } while (pgd
++, addr
= next
, addr
!= end
);
560 static void unmap_hyp_range(pgd_t
*pgdp
, phys_addr_t start
, u64 size
)
562 __unmap_hyp_range(pgdp
, PTRS_PER_PGD
, start
, size
);
565 static void unmap_hyp_idmap_range(pgd_t
*pgdp
, phys_addr_t start
, u64 size
)
567 __unmap_hyp_range(pgdp
, __kvm_idmap_ptrs_per_pgd(), start
, size
);
571 * free_hyp_pgds - free Hyp-mode page tables
573 * Assumes hyp_pgd is a page table used strictly in Hyp-mode and
574 * therefore contains either mappings in the kernel memory area (above
575 * PAGE_OFFSET), or device mappings in the idmap range.
577 * boot_hyp_pgd should only map the idmap range, and is only used in
578 * the extended idmap case.
580 void free_hyp_pgds(void)
584 mutex_lock(&kvm_hyp_pgd_mutex
);
586 id_pgd
= boot_hyp_pgd
? boot_hyp_pgd
: hyp_pgd
;
589 /* In case we never called hyp_mmu_init() */
591 io_map_base
= hyp_idmap_start
;
592 unmap_hyp_idmap_range(id_pgd
, io_map_base
,
593 hyp_idmap_start
+ PAGE_SIZE
- io_map_base
);
597 free_pages((unsigned long)boot_hyp_pgd
, hyp_pgd_order
);
602 unmap_hyp_range(hyp_pgd
, kern_hyp_va(PAGE_OFFSET
),
603 (uintptr_t)high_memory
- PAGE_OFFSET
);
605 free_pages((unsigned long)hyp_pgd
, hyp_pgd_order
);
608 if (merged_hyp_pgd
) {
609 clear_page(merged_hyp_pgd
);
610 free_page((unsigned long)merged_hyp_pgd
);
611 merged_hyp_pgd
= NULL
;
614 mutex_unlock(&kvm_hyp_pgd_mutex
);
617 static void create_hyp_pte_mappings(pmd_t
*pmd
, unsigned long start
,
618 unsigned long end
, unsigned long pfn
,
626 pte
= pte_offset_kernel(pmd
, addr
);
627 kvm_set_pte(pte
, kvm_pfn_pte(pfn
, prot
));
628 get_page(virt_to_page(pte
));
630 } while (addr
+= PAGE_SIZE
, addr
!= end
);
633 static int create_hyp_pmd_mappings(pud_t
*pud
, unsigned long start
,
634 unsigned long end
, unsigned long pfn
,
639 unsigned long addr
, next
;
643 pmd
= pmd_offset(pud
, addr
);
645 BUG_ON(pmd_sect(*pmd
));
647 if (pmd_none(*pmd
)) {
648 pte
= pte_alloc_one_kernel(NULL
);
650 kvm_err("Cannot allocate Hyp pte\n");
653 kvm_pmd_populate(pmd
, pte
);
654 get_page(virt_to_page(pmd
));
657 next
= pmd_addr_end(addr
, end
);
659 create_hyp_pte_mappings(pmd
, addr
, next
, pfn
, prot
);
660 pfn
+= (next
- addr
) >> PAGE_SHIFT
;
661 } while (addr
= next
, addr
!= end
);
666 static int create_hyp_pud_mappings(pgd_t
*pgd
, unsigned long start
,
667 unsigned long end
, unsigned long pfn
,
672 unsigned long addr
, next
;
677 pud
= pud_offset(pgd
, addr
);
679 if (pud_none_or_clear_bad(pud
)) {
680 pmd
= pmd_alloc_one(NULL
, addr
);
682 kvm_err("Cannot allocate Hyp pmd\n");
685 kvm_pud_populate(pud
, pmd
);
686 get_page(virt_to_page(pud
));
689 next
= pud_addr_end(addr
, end
);
690 ret
= create_hyp_pmd_mappings(pud
, addr
, next
, pfn
, prot
);
693 pfn
+= (next
- addr
) >> PAGE_SHIFT
;
694 } while (addr
= next
, addr
!= end
);
699 static int __create_hyp_mappings(pgd_t
*pgdp
, unsigned long ptrs_per_pgd
,
700 unsigned long start
, unsigned long end
,
701 unsigned long pfn
, pgprot_t prot
)
705 unsigned long addr
, next
;
708 mutex_lock(&kvm_hyp_pgd_mutex
);
709 addr
= start
& PAGE_MASK
;
710 end
= PAGE_ALIGN(end
);
712 pgd
= pgdp
+ kvm_pgd_index(addr
, ptrs_per_pgd
);
714 if (pgd_none(*pgd
)) {
715 pud
= pud_alloc_one(NULL
, addr
);
717 kvm_err("Cannot allocate Hyp pud\n");
721 kvm_pgd_populate(pgd
, pud
);
722 get_page(virt_to_page(pgd
));
725 next
= pgd_addr_end(addr
, end
);
726 err
= create_hyp_pud_mappings(pgd
, addr
, next
, pfn
, prot
);
729 pfn
+= (next
- addr
) >> PAGE_SHIFT
;
730 } while (addr
= next
, addr
!= end
);
732 mutex_unlock(&kvm_hyp_pgd_mutex
);
736 static phys_addr_t
kvm_kaddr_to_phys(void *kaddr
)
738 if (!is_vmalloc_addr(kaddr
)) {
739 BUG_ON(!virt_addr_valid(kaddr
));
742 return page_to_phys(vmalloc_to_page(kaddr
)) +
743 offset_in_page(kaddr
);
748 * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
749 * @from: The virtual kernel start address of the range
750 * @to: The virtual kernel end address of the range (exclusive)
751 * @prot: The protection to be applied to this range
753 * The same virtual address as the kernel virtual address is also used
754 * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
757 int create_hyp_mappings(void *from
, void *to
, pgprot_t prot
)
759 phys_addr_t phys_addr
;
760 unsigned long virt_addr
;
761 unsigned long start
= kern_hyp_va((unsigned long)from
);
762 unsigned long end
= kern_hyp_va((unsigned long)to
);
764 if (is_kernel_in_hyp_mode())
767 start
= start
& PAGE_MASK
;
768 end
= PAGE_ALIGN(end
);
770 for (virt_addr
= start
; virt_addr
< end
; virt_addr
+= PAGE_SIZE
) {
773 phys_addr
= kvm_kaddr_to_phys(from
+ virt_addr
- start
);
774 err
= __create_hyp_mappings(hyp_pgd
, PTRS_PER_PGD
,
775 virt_addr
, virt_addr
+ PAGE_SIZE
,
776 __phys_to_pfn(phys_addr
),
785 static int __create_hyp_private_mapping(phys_addr_t phys_addr
, size_t size
,
786 unsigned long *haddr
, pgprot_t prot
)
788 pgd_t
*pgd
= hyp_pgd
;
792 mutex_lock(&kvm_hyp_pgd_mutex
);
795 * This assumes that we we have enough space below the idmap
796 * page to allocate our VAs. If not, the check below will
797 * kick. A potential alternative would be to detect that
798 * overflow and switch to an allocation above the idmap.
800 * The allocated size is always a multiple of PAGE_SIZE.
802 size
= PAGE_ALIGN(size
+ offset_in_page(phys_addr
));
803 base
= io_map_base
- size
;
806 * Verify that BIT(VA_BITS - 1) hasn't been flipped by
807 * allocating the new area, as it would indicate we've
808 * overflowed the idmap/IO address range.
810 if ((base
^ io_map_base
) & BIT(VA_BITS
- 1))
815 mutex_unlock(&kvm_hyp_pgd_mutex
);
820 if (__kvm_cpu_uses_extended_idmap())
823 ret
= __create_hyp_mappings(pgd
, __kvm_idmap_ptrs_per_pgd(),
825 __phys_to_pfn(phys_addr
), prot
);
829 *haddr
= base
+ offset_in_page(phys_addr
);
836 * create_hyp_io_mappings - Map IO into both kernel and HYP
837 * @phys_addr: The physical start address which gets mapped
838 * @size: Size of the region being mapped
839 * @kaddr: Kernel VA for this mapping
840 * @haddr: HYP VA for this mapping
842 int create_hyp_io_mappings(phys_addr_t phys_addr
, size_t size
,
843 void __iomem
**kaddr
,
844 void __iomem
**haddr
)
849 *kaddr
= ioremap(phys_addr
, size
);
853 if (is_kernel_in_hyp_mode()) {
858 ret
= __create_hyp_private_mapping(phys_addr
, size
,
859 &addr
, PAGE_HYP_DEVICE
);
867 *haddr
= (void __iomem
*)addr
;
872 * create_hyp_exec_mappings - Map an executable range into HYP
873 * @phys_addr: The physical start address which gets mapped
874 * @size: Size of the region being mapped
875 * @haddr: HYP VA for this mapping
877 int create_hyp_exec_mappings(phys_addr_t phys_addr
, size_t size
,
883 BUG_ON(is_kernel_in_hyp_mode());
885 ret
= __create_hyp_private_mapping(phys_addr
, size
,
886 &addr
, PAGE_HYP_EXEC
);
892 *haddr
= (void *)addr
;
897 * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
898 * @kvm: The KVM struct pointer for the VM.
900 * Allocates only the stage-2 HW PGD level table(s) of size defined by
901 * stage2_pgd_size(kvm).
903 * Note we don't need locking here as this is only called when the VM is
904 * created, which can only be done once.
906 int kvm_alloc_stage2_pgd(struct kvm
*kvm
)
908 phys_addr_t pgd_phys
;
911 if (kvm
->arch
.pgd
!= NULL
) {
912 kvm_err("kvm_arch already initialized?\n");
916 /* Allocate the HW PGD, making sure that each page gets its own refcount */
917 pgd
= alloc_pages_exact(stage2_pgd_size(kvm
), GFP_KERNEL
| __GFP_ZERO
);
921 pgd_phys
= virt_to_phys(pgd
);
922 if (WARN_ON(pgd_phys
& ~kvm_vttbr_baddr_mask(kvm
)))
926 kvm
->arch
.pgd_phys
= pgd_phys
;
930 static void stage2_unmap_memslot(struct kvm
*kvm
,
931 struct kvm_memory_slot
*memslot
)
933 hva_t hva
= memslot
->userspace_addr
;
934 phys_addr_t addr
= memslot
->base_gfn
<< PAGE_SHIFT
;
935 phys_addr_t size
= PAGE_SIZE
* memslot
->npages
;
936 hva_t reg_end
= hva
+ size
;
939 * A memory region could potentially cover multiple VMAs, and any holes
940 * between them, so iterate over all of them to find out if we should
943 * +--------------------------------------------+
944 * +---------------+----------------+ +----------------+
945 * | : VMA 1 | VMA 2 | | VMA 3 : |
946 * +---------------+----------------+ +----------------+
948 * +--------------------------------------------+
951 struct vm_area_struct
*vma
= find_vma(current
->mm
, hva
);
952 hva_t vm_start
, vm_end
;
954 if (!vma
|| vma
->vm_start
>= reg_end
)
958 * Take the intersection of this VMA with the memory region
960 vm_start
= max(hva
, vma
->vm_start
);
961 vm_end
= min(reg_end
, vma
->vm_end
);
963 if (!(vma
->vm_flags
& VM_PFNMAP
)) {
964 gpa_t gpa
= addr
+ (vm_start
- memslot
->userspace_addr
);
965 unmap_stage2_range(kvm
, gpa
, vm_end
- vm_start
);
968 } while (hva
< reg_end
);
972 * stage2_unmap_vm - Unmap Stage-2 RAM mappings
973 * @kvm: The struct kvm pointer
975 * Go through the memregions and unmap any reguler RAM
976 * backing memory already mapped to the VM.
978 void stage2_unmap_vm(struct kvm
*kvm
)
980 struct kvm_memslots
*slots
;
981 struct kvm_memory_slot
*memslot
;
984 idx
= srcu_read_lock(&kvm
->srcu
);
985 down_read(¤t
->mm
->mmap_sem
);
986 spin_lock(&kvm
->mmu_lock
);
988 slots
= kvm_memslots(kvm
);
989 kvm_for_each_memslot(memslot
, slots
)
990 stage2_unmap_memslot(kvm
, memslot
);
992 spin_unlock(&kvm
->mmu_lock
);
993 up_read(¤t
->mm
->mmap_sem
);
994 srcu_read_unlock(&kvm
->srcu
, idx
);
998 * kvm_free_stage2_pgd - free all stage-2 tables
999 * @kvm: The KVM struct pointer for the VM.
1001 * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
1002 * underlying level-2 and level-3 tables before freeing the actual level-1 table
1003 * and setting the struct pointer to NULL.
1005 void kvm_free_stage2_pgd(struct kvm
*kvm
)
1009 spin_lock(&kvm
->mmu_lock
);
1010 if (kvm
->arch
.pgd
) {
1011 unmap_stage2_range(kvm
, 0, kvm_phys_size(kvm
));
1012 pgd
= READ_ONCE(kvm
->arch
.pgd
);
1013 kvm
->arch
.pgd
= NULL
;
1014 kvm
->arch
.pgd_phys
= 0;
1016 spin_unlock(&kvm
->mmu_lock
);
1018 /* Free the HW pgd, one page at a time */
1020 free_pages_exact(pgd
, stage2_pgd_size(kvm
));
1023 static pud_t
*stage2_get_pud(struct kvm
*kvm
, struct kvm_mmu_memory_cache
*cache
,
1029 pgd
= kvm
->arch
.pgd
+ stage2_pgd_index(kvm
, addr
);
1030 if (stage2_pgd_none(kvm
, *pgd
)) {
1033 pud
= mmu_memory_cache_alloc(cache
);
1034 stage2_pgd_populate(kvm
, pgd
, pud
);
1035 get_page(virt_to_page(pgd
));
1038 return stage2_pud_offset(kvm
, pgd
, addr
);
1041 static pmd_t
*stage2_get_pmd(struct kvm
*kvm
, struct kvm_mmu_memory_cache
*cache
,
1047 pud
= stage2_get_pud(kvm
, cache
, addr
);
1048 if (!pud
|| stage2_pud_huge(kvm
, *pud
))
1051 if (stage2_pud_none(kvm
, *pud
)) {
1054 pmd
= mmu_memory_cache_alloc(cache
);
1055 stage2_pud_populate(kvm
, pud
, pmd
);
1056 get_page(virt_to_page(pud
));
1059 return stage2_pmd_offset(kvm
, pud
, addr
);
1062 static int stage2_set_pmd_huge(struct kvm
*kvm
, struct kvm_mmu_memory_cache
1063 *cache
, phys_addr_t addr
, const pmd_t
*new_pmd
)
1065 pmd_t
*pmd
, old_pmd
;
1068 pmd
= stage2_get_pmd(kvm
, cache
, addr
);
1073 * Multiple vcpus faulting on the same PMD entry, can
1074 * lead to them sequentially updating the PMD with the
1075 * same value. Following the break-before-make
1076 * (pmd_clear() followed by tlb_flush()) process can
1077 * hinder forward progress due to refaults generated
1078 * on missing translations.
1080 * Skip updating the page table if the entry is
1083 if (pmd_val(old_pmd
) == pmd_val(*new_pmd
))
1086 if (pmd_present(old_pmd
)) {
1088 * If we already have PTE level mapping for this block,
1089 * we must unmap it to avoid inconsistent TLB state and
1090 * leaking the table page. We could end up in this situation
1091 * if the memory slot was marked for dirty logging and was
1092 * reverted, leaving PTE level mappings for the pages accessed
1093 * during the period. So, unmap the PTE level mapping for this
1094 * block and retry, as we could have released the upper level
1095 * table in the process.
1097 * Normal THP split/merge follows mmu_notifier callbacks and do
1098 * get handled accordingly.
1100 if (!pmd_thp_or_huge(old_pmd
)) {
1101 unmap_stage2_range(kvm
, addr
& S2_PMD_MASK
, S2_PMD_SIZE
);
1105 * Mapping in huge pages should only happen through a
1106 * fault. If a page is merged into a transparent huge
1107 * page, the individual subpages of that huge page
1108 * should be unmapped through MMU notifiers before we
1111 * Merging of CompoundPages is not supported; they
1112 * should become splitting first, unmapped, merged,
1113 * and mapped back in on-demand.
1115 WARN_ON_ONCE(pmd_pfn(old_pmd
) != pmd_pfn(*new_pmd
));
1117 kvm_tlb_flush_vmid_ipa(kvm
, addr
);
1119 get_page(virt_to_page(pmd
));
1122 kvm_set_pmd(pmd
, *new_pmd
);
1126 static int stage2_set_pud_huge(struct kvm
*kvm
, struct kvm_mmu_memory_cache
*cache
,
1127 phys_addr_t addr
, const pud_t
*new_pudp
)
1129 pud_t
*pudp
, old_pud
;
1132 pudp
= stage2_get_pud(kvm
, cache
, addr
);
1138 * A large number of vcpus faulting on the same stage 2 entry,
1139 * can lead to a refault due to the stage2_pud_clear()/tlb_flush().
1140 * Skip updating the page tables if there is no change.
1142 if (pud_val(old_pud
) == pud_val(*new_pudp
))
1145 if (stage2_pud_present(kvm
, old_pud
)) {
1147 * If we already have table level mapping for this block, unmap
1148 * the range for this block and retry.
1150 if (!stage2_pud_huge(kvm
, old_pud
)) {
1151 unmap_stage2_range(kvm
, addr
& S2_PUD_MASK
, S2_PUD_SIZE
);
1155 WARN_ON_ONCE(kvm_pud_pfn(old_pud
) != kvm_pud_pfn(*new_pudp
));
1156 stage2_pud_clear(kvm
, pudp
);
1157 kvm_tlb_flush_vmid_ipa(kvm
, addr
);
1159 get_page(virt_to_page(pudp
));
1162 kvm_set_pud(pudp
, *new_pudp
);
1167 * stage2_get_leaf_entry - walk the stage2 VM page tables and return
1168 * true if a valid and present leaf-entry is found. A pointer to the
1169 * leaf-entry is returned in the appropriate level variable - pudpp,
1172 static bool stage2_get_leaf_entry(struct kvm
*kvm
, phys_addr_t addr
,
1173 pud_t
**pudpp
, pmd_t
**pmdpp
, pte_t
**ptepp
)
1183 pudp
= stage2_get_pud(kvm
, NULL
, addr
);
1184 if (!pudp
|| stage2_pud_none(kvm
, *pudp
) || !stage2_pud_present(kvm
, *pudp
))
1187 if (stage2_pud_huge(kvm
, *pudp
)) {
1192 pmdp
= stage2_pmd_offset(kvm
, pudp
, addr
);
1193 if (!pmdp
|| pmd_none(*pmdp
) || !pmd_present(*pmdp
))
1196 if (pmd_thp_or_huge(*pmdp
)) {
1201 ptep
= pte_offset_kernel(pmdp
, addr
);
1202 if (!ptep
|| pte_none(*ptep
) || !pte_present(*ptep
))
1209 static bool stage2_is_exec(struct kvm
*kvm
, phys_addr_t addr
)
1216 found
= stage2_get_leaf_entry(kvm
, addr
, &pudp
, &pmdp
, &ptep
);
1221 return kvm_s2pud_exec(pudp
);
1223 return kvm_s2pmd_exec(pmdp
);
1225 return kvm_s2pte_exec(ptep
);
1228 static int stage2_set_pte(struct kvm
*kvm
, struct kvm_mmu_memory_cache
*cache
,
1229 phys_addr_t addr
, const pte_t
*new_pte
,
1230 unsigned long flags
)
1234 pte_t
*pte
, old_pte
;
1235 bool iomap
= flags
& KVM_S2PTE_FLAG_IS_IOMAP
;
1236 bool logging_active
= flags
& KVM_S2_FLAG_LOGGING_ACTIVE
;
1238 VM_BUG_ON(logging_active
&& !cache
);
1240 /* Create stage-2 page table mapping - Levels 0 and 1 */
1241 pud
= stage2_get_pud(kvm
, cache
, addr
);
1244 * Ignore calls from kvm_set_spte_hva for unallocated
1251 * While dirty page logging - dissolve huge PUD, then continue
1252 * on to allocate page.
1255 stage2_dissolve_pud(kvm
, addr
, pud
);
1257 if (stage2_pud_none(kvm
, *pud
)) {
1259 return 0; /* ignore calls from kvm_set_spte_hva */
1260 pmd
= mmu_memory_cache_alloc(cache
);
1261 stage2_pud_populate(kvm
, pud
, pmd
);
1262 get_page(virt_to_page(pud
));
1265 pmd
= stage2_pmd_offset(kvm
, pud
, addr
);
1268 * Ignore calls from kvm_set_spte_hva for unallocated
1275 * While dirty page logging - dissolve huge PMD, then continue on to
1279 stage2_dissolve_pmd(kvm
, addr
, pmd
);
1281 /* Create stage-2 page mappings - Level 2 */
1282 if (pmd_none(*pmd
)) {
1284 return 0; /* ignore calls from kvm_set_spte_hva */
1285 pte
= mmu_memory_cache_alloc(cache
);
1286 kvm_pmd_populate(pmd
, pte
);
1287 get_page(virt_to_page(pmd
));
1290 pte
= pte_offset_kernel(pmd
, addr
);
1292 if (iomap
&& pte_present(*pte
))
1295 /* Create 2nd stage page table mapping - Level 3 */
1297 if (pte_present(old_pte
)) {
1298 /* Skip page table update if there is no change */
1299 if (pte_val(old_pte
) == pte_val(*new_pte
))
1302 kvm_set_pte(pte
, __pte(0));
1303 kvm_tlb_flush_vmid_ipa(kvm
, addr
);
1305 get_page(virt_to_page(pte
));
1308 kvm_set_pte(pte
, *new_pte
);
1312 #ifndef __HAVE_ARCH_PTEP_TEST_AND_CLEAR_YOUNG
1313 static int stage2_ptep_test_and_clear_young(pte_t
*pte
)
1315 if (pte_young(*pte
)) {
1316 *pte
= pte_mkold(*pte
);
1322 static int stage2_ptep_test_and_clear_young(pte_t
*pte
)
1324 return __ptep_test_and_clear_young(pte
);
1328 static int stage2_pmdp_test_and_clear_young(pmd_t
*pmd
)
1330 return stage2_ptep_test_and_clear_young((pte_t
*)pmd
);
1333 static int stage2_pudp_test_and_clear_young(pud_t
*pud
)
1335 return stage2_ptep_test_and_clear_young((pte_t
*)pud
);
1339 * kvm_phys_addr_ioremap - map a device range to guest IPA
1341 * @kvm: The KVM pointer
1342 * @guest_ipa: The IPA at which to insert the mapping
1343 * @pa: The physical address of the device
1344 * @size: The size of the mapping
1346 int kvm_phys_addr_ioremap(struct kvm
*kvm
, phys_addr_t guest_ipa
,
1347 phys_addr_t pa
, unsigned long size
, bool writable
)
1349 phys_addr_t addr
, end
;
1352 struct kvm_mmu_memory_cache cache
= { 0, };
1354 end
= (guest_ipa
+ size
+ PAGE_SIZE
- 1) & PAGE_MASK
;
1355 pfn
= __phys_to_pfn(pa
);
1357 for (addr
= guest_ipa
; addr
< end
; addr
+= PAGE_SIZE
) {
1358 pte_t pte
= kvm_pfn_pte(pfn
, PAGE_S2_DEVICE
);
1361 pte
= kvm_s2pte_mkwrite(pte
);
1363 ret
= mmu_topup_memory_cache(&cache
,
1364 kvm_mmu_cache_min_pages(kvm
),
1368 spin_lock(&kvm
->mmu_lock
);
1369 ret
= stage2_set_pte(kvm
, &cache
, addr
, &pte
,
1370 KVM_S2PTE_FLAG_IS_IOMAP
);
1371 spin_unlock(&kvm
->mmu_lock
);
1379 mmu_free_memory_cache(&cache
);
1383 static bool transparent_hugepage_adjust(kvm_pfn_t
*pfnp
, phys_addr_t
*ipap
)
1385 kvm_pfn_t pfn
= *pfnp
;
1386 gfn_t gfn
= *ipap
>> PAGE_SHIFT
;
1387 struct page
*page
= pfn_to_page(pfn
);
1390 * PageTransCompoundMap() returns true for THP and
1391 * hugetlbfs. Make sure the adjustment is done only for THP
1394 if (!PageHuge(page
) && PageTransCompoundMap(page
)) {
1397 * The address we faulted on is backed by a transparent huge
1398 * page. However, because we map the compound huge page and
1399 * not the individual tail page, we need to transfer the
1400 * refcount to the head page. We have to be careful that the
1401 * THP doesn't start to split while we are adjusting the
1404 * We are sure this doesn't happen, because mmu_notifier_retry
1405 * was successful and we are holding the mmu_lock, so if this
1406 * THP is trying to split, it will be blocked in the mmu
1407 * notifier before touching any of the pages, specifically
1408 * before being able to call __split_huge_page_refcount().
1410 * We can therefore safely transfer the refcount from PG_tail
1411 * to PG_head and switch the pfn from a tail page to the head
1414 mask
= PTRS_PER_PMD
- 1;
1415 VM_BUG_ON((gfn
& mask
) != (pfn
& mask
));
1418 kvm_release_pfn_clean(pfn
);
1431 * stage2_wp_ptes - write protect PMD range
1432 * @pmd: pointer to pmd entry
1433 * @addr: range start address
1434 * @end: range end address
1436 static void stage2_wp_ptes(pmd_t
*pmd
, phys_addr_t addr
, phys_addr_t end
)
1440 pte
= pte_offset_kernel(pmd
, addr
);
1442 if (!pte_none(*pte
)) {
1443 if (!kvm_s2pte_readonly(pte
))
1444 kvm_set_s2pte_readonly(pte
);
1446 } while (pte
++, addr
+= PAGE_SIZE
, addr
!= end
);
1450 * stage2_wp_pmds - write protect PUD range
1451 * kvm: kvm instance for the VM
1452 * @pud: pointer to pud entry
1453 * @addr: range start address
1454 * @end: range end address
1456 static void stage2_wp_pmds(struct kvm
*kvm
, pud_t
*pud
,
1457 phys_addr_t addr
, phys_addr_t end
)
1462 pmd
= stage2_pmd_offset(kvm
, pud
, addr
);
1465 next
= stage2_pmd_addr_end(kvm
, addr
, end
);
1466 if (!pmd_none(*pmd
)) {
1467 if (pmd_thp_or_huge(*pmd
)) {
1468 if (!kvm_s2pmd_readonly(pmd
))
1469 kvm_set_s2pmd_readonly(pmd
);
1471 stage2_wp_ptes(pmd
, addr
, next
);
1474 } while (pmd
++, addr
= next
, addr
!= end
);
1478 * stage2_wp_puds - write protect PGD range
1479 * @pgd: pointer to pgd entry
1480 * @addr: range start address
1481 * @end: range end address
1483 static void stage2_wp_puds(struct kvm
*kvm
, pgd_t
*pgd
,
1484 phys_addr_t addr
, phys_addr_t end
)
1489 pud
= stage2_pud_offset(kvm
, pgd
, addr
);
1491 next
= stage2_pud_addr_end(kvm
, addr
, end
);
1492 if (!stage2_pud_none(kvm
, *pud
)) {
1493 if (stage2_pud_huge(kvm
, *pud
)) {
1494 if (!kvm_s2pud_readonly(pud
))
1495 kvm_set_s2pud_readonly(pud
);
1497 stage2_wp_pmds(kvm
, pud
, addr
, next
);
1500 } while (pud
++, addr
= next
, addr
!= end
);
1504 * stage2_wp_range() - write protect stage2 memory region range
1505 * @kvm: The KVM pointer
1506 * @addr: Start address of range
1507 * @end: End address of range
1509 static void stage2_wp_range(struct kvm
*kvm
, phys_addr_t addr
, phys_addr_t end
)
1514 pgd
= kvm
->arch
.pgd
+ stage2_pgd_index(kvm
, addr
);
1517 * Release kvm_mmu_lock periodically if the memory region is
1518 * large. Otherwise, we may see kernel panics with
1519 * CONFIG_DETECT_HUNG_TASK, CONFIG_LOCKUP_DETECTOR,
1520 * CONFIG_LOCKDEP. Additionally, holding the lock too long
1521 * will also starve other vCPUs. We have to also make sure
1522 * that the page tables are not freed while we released
1525 cond_resched_lock(&kvm
->mmu_lock
);
1526 if (!READ_ONCE(kvm
->arch
.pgd
))
1528 next
= stage2_pgd_addr_end(kvm
, addr
, end
);
1529 if (stage2_pgd_present(kvm
, *pgd
))
1530 stage2_wp_puds(kvm
, pgd
, addr
, next
);
1531 } while (pgd
++, addr
= next
, addr
!= end
);
1535 * kvm_mmu_wp_memory_region() - write protect stage 2 entries for memory slot
1536 * @kvm: The KVM pointer
1537 * @slot: The memory slot to write protect
1539 * Called to start logging dirty pages after memory region
1540 * KVM_MEM_LOG_DIRTY_PAGES operation is called. After this function returns
1541 * all present PUD, PMD and PTEs are write protected in the memory region.
1542 * Afterwards read of dirty page log can be called.
1544 * Acquires kvm_mmu_lock. Called with kvm->slots_lock mutex acquired,
1545 * serializing operations for VM memory regions.
1547 void kvm_mmu_wp_memory_region(struct kvm
*kvm
, int slot
)
1549 struct kvm_memslots
*slots
= kvm_memslots(kvm
);
1550 struct kvm_memory_slot
*memslot
= id_to_memslot(slots
, slot
);
1551 phys_addr_t start
= memslot
->base_gfn
<< PAGE_SHIFT
;
1552 phys_addr_t end
= (memslot
->base_gfn
+ memslot
->npages
) << PAGE_SHIFT
;
1554 spin_lock(&kvm
->mmu_lock
);
1555 stage2_wp_range(kvm
, start
, end
);
1556 spin_unlock(&kvm
->mmu_lock
);
1557 kvm_flush_remote_tlbs(kvm
);
1561 * kvm_mmu_write_protect_pt_masked() - write protect dirty pages
1562 * @kvm: The KVM pointer
1563 * @slot: The memory slot associated with mask
1564 * @gfn_offset: The gfn offset in memory slot
1565 * @mask: The mask of dirty pages at offset 'gfn_offset' in this memory
1566 * slot to be write protected
1568 * Walks bits set in mask write protects the associated pte's. Caller must
1569 * acquire kvm_mmu_lock.
1571 static void kvm_mmu_write_protect_pt_masked(struct kvm
*kvm
,
1572 struct kvm_memory_slot
*slot
,
1573 gfn_t gfn_offset
, unsigned long mask
)
1575 phys_addr_t base_gfn
= slot
->base_gfn
+ gfn_offset
;
1576 phys_addr_t start
= (base_gfn
+ __ffs(mask
)) << PAGE_SHIFT
;
1577 phys_addr_t end
= (base_gfn
+ __fls(mask
) + 1) << PAGE_SHIFT
;
1579 stage2_wp_range(kvm
, start
, end
);
1583 * kvm_arch_mmu_enable_log_dirty_pt_masked - enable dirty logging for selected
1586 * It calls kvm_mmu_write_protect_pt_masked to write protect selected pages to
1587 * enable dirty logging for them.
1589 void kvm_arch_mmu_enable_log_dirty_pt_masked(struct kvm
*kvm
,
1590 struct kvm_memory_slot
*slot
,
1591 gfn_t gfn_offset
, unsigned long mask
)
1593 kvm_mmu_write_protect_pt_masked(kvm
, slot
, gfn_offset
, mask
);
1596 static void clean_dcache_guest_page(kvm_pfn_t pfn
, unsigned long size
)
1598 __clean_dcache_guest_page(pfn
, size
);
1601 static void invalidate_icache_guest_page(kvm_pfn_t pfn
, unsigned long size
)
1603 __invalidate_icache_guest_page(pfn
, size
);
1606 static void kvm_send_hwpoison_signal(unsigned long address
,
1607 struct vm_area_struct
*vma
)
1611 if (is_vm_hugetlb_page(vma
))
1612 lsb
= huge_page_shift(hstate_vma(vma
));
1616 send_sig_mceerr(BUS_MCEERR_AR
, (void __user
*)address
, lsb
, current
);
1619 static bool fault_supports_stage2_huge_mapping(struct kvm_memory_slot
*memslot
,
1621 unsigned long map_size
)
1624 hva_t uaddr_start
, uaddr_end
;
1627 size
= memslot
->npages
* PAGE_SIZE
;
1629 gpa_start
= memslot
->base_gfn
<< PAGE_SHIFT
;
1631 uaddr_start
= memslot
->userspace_addr
;
1632 uaddr_end
= uaddr_start
+ size
;
1635 * Pages belonging to memslots that don't have the same alignment
1636 * within a PMD/PUD for userspace and IPA cannot be mapped with stage-2
1637 * PMD/PUD entries, because we'll end up mapping the wrong pages.
1639 * Consider a layout like the following:
1641 * memslot->userspace_addr:
1642 * +-----+--------------------+--------------------+---+
1643 * |abcde|fgh Stage-1 block | Stage-1 block tv|xyz|
1644 * +-----+--------------------+--------------------+---+
1646 * memslot->base_gfn << PAGE_SIZE:
1647 * +---+--------------------+--------------------+-----+
1648 * |abc|def Stage-2 block | Stage-2 block |tvxyz|
1649 * +---+--------------------+--------------------+-----+
1651 * If we create those stage-2 blocks, we'll end up with this incorrect
1657 if ((gpa_start
& (map_size
- 1)) != (uaddr_start
& (map_size
- 1)))
1661 * Next, let's make sure we're not trying to map anything not covered
1662 * by the memslot. This means we have to prohibit block size mappings
1663 * for the beginning and end of a non-block aligned and non-block sized
1664 * memory slot (illustrated by the head and tail parts of the
1665 * userspace view above containing pages 'abcde' and 'xyz',
1668 * Note that it doesn't matter if we do the check using the
1669 * userspace_addr or the base_gfn, as both are equally aligned (per
1670 * the check above) and equally sized.
1672 return (hva
& ~(map_size
- 1)) >= uaddr_start
&&
1673 (hva
& ~(map_size
- 1)) + map_size
<= uaddr_end
;
1676 static int user_mem_abort(struct kvm_vcpu
*vcpu
, phys_addr_t fault_ipa
,
1677 struct kvm_memory_slot
*memslot
, unsigned long hva
,
1678 unsigned long fault_status
)
1681 bool write_fault
, writable
, force_pte
= false;
1682 bool exec_fault
, needs_exec
;
1683 unsigned long mmu_seq
;
1684 gfn_t gfn
= fault_ipa
>> PAGE_SHIFT
;
1685 struct kvm
*kvm
= vcpu
->kvm
;
1686 struct kvm_mmu_memory_cache
*memcache
= &vcpu
->arch
.mmu_page_cache
;
1687 struct vm_area_struct
*vma
;
1689 pgprot_t mem_type
= PAGE_S2
;
1690 bool logging_active
= memslot_is_logging(memslot
);
1691 unsigned long vma_pagesize
, flags
= 0;
1693 write_fault
= kvm_is_write_fault(vcpu
);
1694 exec_fault
= kvm_vcpu_trap_is_iabt(vcpu
);
1695 VM_BUG_ON(write_fault
&& exec_fault
);
1697 if (fault_status
== FSC_PERM
&& !write_fault
&& !exec_fault
) {
1698 kvm_err("Unexpected L2 read permission error\n");
1702 /* Let's check if we will get back a huge page backed by hugetlbfs */
1703 down_read(¤t
->mm
->mmap_sem
);
1704 vma
= find_vma_intersection(current
->mm
, hva
, hva
+ 1);
1705 if (unlikely(!vma
)) {
1706 kvm_err("Failed to find VMA for hva 0x%lx\n", hva
);
1707 up_read(¤t
->mm
->mmap_sem
);
1711 vma_pagesize
= vma_kernel_pagesize(vma
);
1712 if (logging_active
||
1713 !fault_supports_stage2_huge_mapping(memslot
, hva
, vma_pagesize
)) {
1715 vma_pagesize
= PAGE_SIZE
;
1719 * The stage2 has a minimum of 2 level table (For arm64 see
1720 * kvm_arm_setup_stage2()). Hence, we are guaranteed that we can
1721 * use PMD_SIZE huge mappings (even when the PMD is folded into PGD).
1722 * As for PUD huge maps, we must make sure that we have at least
1723 * 3 levels, i.e, PMD is not folded.
1725 if (vma_pagesize
== PMD_SIZE
||
1726 (vma_pagesize
== PUD_SIZE
&& kvm_stage2_has_pmd(kvm
)))
1727 gfn
= (fault_ipa
& huge_page_mask(hstate_vma(vma
))) >> PAGE_SHIFT
;
1728 up_read(¤t
->mm
->mmap_sem
);
1730 /* We need minimum second+third level pages */
1731 ret
= mmu_topup_memory_cache(memcache
, kvm_mmu_cache_min_pages(kvm
),
1736 mmu_seq
= vcpu
->kvm
->mmu_notifier_seq
;
1738 * Ensure the read of mmu_notifier_seq happens before we call
1739 * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
1740 * the page we just got a reference to gets unmapped before we have a
1741 * chance to grab the mmu_lock, which ensure that if the page gets
1742 * unmapped afterwards, the call to kvm_unmap_hva will take it away
1743 * from us again properly. This smp_rmb() interacts with the smp_wmb()
1744 * in kvm_mmu_notifier_invalidate_<page|range_end>.
1748 pfn
= gfn_to_pfn_prot(kvm
, gfn
, write_fault
, &writable
);
1749 if (pfn
== KVM_PFN_ERR_HWPOISON
) {
1750 kvm_send_hwpoison_signal(hva
, vma
);
1753 if (is_error_noslot_pfn(pfn
))
1756 if (kvm_is_device_pfn(pfn
)) {
1757 mem_type
= PAGE_S2_DEVICE
;
1758 flags
|= KVM_S2PTE_FLAG_IS_IOMAP
;
1759 } else if (logging_active
) {
1761 * Faults on pages in a memslot with logging enabled
1762 * should not be mapped with huge pages (it introduces churn
1763 * and performance degradation), so force a pte mapping.
1765 flags
|= KVM_S2_FLAG_LOGGING_ACTIVE
;
1768 * Only actually map the page as writable if this was a write
1775 spin_lock(&kvm
->mmu_lock
);
1776 if (mmu_notifier_retry(kvm
, mmu_seq
))
1779 if (vma_pagesize
== PAGE_SIZE
&& !force_pte
) {
1781 * Only PMD_SIZE transparent hugepages(THP) are
1782 * currently supported. This code will need to be
1783 * updated to support other THP sizes.
1785 * Make sure the host VA and the guest IPA are sufficiently
1786 * aligned and that the block is contained within the memslot.
1788 if (fault_supports_stage2_huge_mapping(memslot
, hva
, PMD_SIZE
) &&
1789 transparent_hugepage_adjust(&pfn
, &fault_ipa
))
1790 vma_pagesize
= PMD_SIZE
;
1794 kvm_set_pfn_dirty(pfn
);
1796 if (fault_status
!= FSC_PERM
)
1797 clean_dcache_guest_page(pfn
, vma_pagesize
);
1800 invalidate_icache_guest_page(pfn
, vma_pagesize
);
1803 * If we took an execution fault we have made the
1804 * icache/dcache coherent above and should now let the s2
1805 * mapping be executable.
1807 * Write faults (!exec_fault && FSC_PERM) are orthogonal to
1808 * execute permissions, and we preserve whatever we have.
1810 needs_exec
= exec_fault
||
1811 (fault_status
== FSC_PERM
&& stage2_is_exec(kvm
, fault_ipa
));
1813 if (vma_pagesize
== PUD_SIZE
) {
1814 pud_t new_pud
= kvm_pfn_pud(pfn
, mem_type
);
1816 new_pud
= kvm_pud_mkhuge(new_pud
);
1818 new_pud
= kvm_s2pud_mkwrite(new_pud
);
1821 new_pud
= kvm_s2pud_mkexec(new_pud
);
1823 ret
= stage2_set_pud_huge(kvm
, memcache
, fault_ipa
, &new_pud
);
1824 } else if (vma_pagesize
== PMD_SIZE
) {
1825 pmd_t new_pmd
= kvm_pfn_pmd(pfn
, mem_type
);
1827 new_pmd
= kvm_pmd_mkhuge(new_pmd
);
1830 new_pmd
= kvm_s2pmd_mkwrite(new_pmd
);
1833 new_pmd
= kvm_s2pmd_mkexec(new_pmd
);
1835 ret
= stage2_set_pmd_huge(kvm
, memcache
, fault_ipa
, &new_pmd
);
1837 pte_t new_pte
= kvm_pfn_pte(pfn
, mem_type
);
1840 new_pte
= kvm_s2pte_mkwrite(new_pte
);
1841 mark_page_dirty(kvm
, gfn
);
1845 new_pte
= kvm_s2pte_mkexec(new_pte
);
1847 ret
= stage2_set_pte(kvm
, memcache
, fault_ipa
, &new_pte
, flags
);
1851 spin_unlock(&kvm
->mmu_lock
);
1852 kvm_set_pfn_accessed(pfn
);
1853 kvm_release_pfn_clean(pfn
);
1858 * Resolve the access fault by making the page young again.
1859 * Note that because the faulting entry is guaranteed not to be
1860 * cached in the TLB, we don't need to invalidate anything.
1861 * Only the HW Access Flag updates are supported for Stage 2 (no DBM),
1862 * so there is no need for atomic (pte|pmd)_mkyoung operations.
1864 static void handle_access_fault(struct kvm_vcpu
*vcpu
, phys_addr_t fault_ipa
)
1870 bool pfn_valid
= false;
1872 trace_kvm_access_fault(fault_ipa
);
1874 spin_lock(&vcpu
->kvm
->mmu_lock
);
1876 if (!stage2_get_leaf_entry(vcpu
->kvm
, fault_ipa
, &pud
, &pmd
, &pte
))
1879 if (pud
) { /* HugeTLB */
1880 *pud
= kvm_s2pud_mkyoung(*pud
);
1881 pfn
= kvm_pud_pfn(*pud
);
1883 } else if (pmd
) { /* THP, HugeTLB */
1884 *pmd
= pmd_mkyoung(*pmd
);
1885 pfn
= pmd_pfn(*pmd
);
1888 *pte
= pte_mkyoung(*pte
); /* Just a page... */
1889 pfn
= pte_pfn(*pte
);
1894 spin_unlock(&vcpu
->kvm
->mmu_lock
);
1896 kvm_set_pfn_accessed(pfn
);
1900 * kvm_handle_guest_abort - handles all 2nd stage aborts
1901 * @vcpu: the VCPU pointer
1902 * @run: the kvm_run structure
1904 * Any abort that gets to the host is almost guaranteed to be caused by a
1905 * missing second stage translation table entry, which can mean that either the
1906 * guest simply needs more memory and we must allocate an appropriate page or it
1907 * can mean that the guest tried to access I/O memory, which is emulated by user
1908 * space. The distinction is based on the IPA causing the fault and whether this
1909 * memory region has been registered as standard RAM by user space.
1911 int kvm_handle_guest_abort(struct kvm_vcpu
*vcpu
, struct kvm_run
*run
)
1913 unsigned long fault_status
;
1914 phys_addr_t fault_ipa
;
1915 struct kvm_memory_slot
*memslot
;
1917 bool is_iabt
, write_fault
, writable
;
1921 fault_status
= kvm_vcpu_trap_get_fault_type(vcpu
);
1923 fault_ipa
= kvm_vcpu_get_fault_ipa(vcpu
);
1924 is_iabt
= kvm_vcpu_trap_is_iabt(vcpu
);
1926 /* Synchronous External Abort? */
1927 if (kvm_vcpu_dabt_isextabt(vcpu
)) {
1929 * For RAS the host kernel may handle this abort.
1930 * There is no need to pass the error into the guest.
1932 if (!kvm_handle_guest_sea(fault_ipa
, kvm_vcpu_get_hsr(vcpu
)))
1935 if (unlikely(!is_iabt
)) {
1936 kvm_inject_vabt(vcpu
);
1941 trace_kvm_guest_fault(*vcpu_pc(vcpu
), kvm_vcpu_get_hsr(vcpu
),
1942 kvm_vcpu_get_hfar(vcpu
), fault_ipa
);
1944 /* Check the stage-2 fault is trans. fault or write fault */
1945 if (fault_status
!= FSC_FAULT
&& fault_status
!= FSC_PERM
&&
1946 fault_status
!= FSC_ACCESS
) {
1947 kvm_err("Unsupported FSC: EC=%#x xFSC=%#lx ESR_EL2=%#lx\n",
1948 kvm_vcpu_trap_get_class(vcpu
),
1949 (unsigned long)kvm_vcpu_trap_get_fault(vcpu
),
1950 (unsigned long)kvm_vcpu_get_hsr(vcpu
));
1954 idx
= srcu_read_lock(&vcpu
->kvm
->srcu
);
1956 gfn
= fault_ipa
>> PAGE_SHIFT
;
1957 memslot
= gfn_to_memslot(vcpu
->kvm
, gfn
);
1958 hva
= gfn_to_hva_memslot_prot(memslot
, gfn
, &writable
);
1959 write_fault
= kvm_is_write_fault(vcpu
);
1960 if (kvm_is_error_hva(hva
) || (write_fault
&& !writable
)) {
1962 /* Prefetch Abort on I/O address */
1963 kvm_inject_pabt(vcpu
, kvm_vcpu_get_hfar(vcpu
));
1969 * Check for a cache maintenance operation. Since we
1970 * ended-up here, we know it is outside of any memory
1971 * slot. But we can't find out if that is for a device,
1972 * or if the guest is just being stupid. The only thing
1973 * we know for sure is that this range cannot be cached.
1975 * So let's assume that the guest is just being
1976 * cautious, and skip the instruction.
1978 if (kvm_vcpu_dabt_is_cm(vcpu
)) {
1979 kvm_skip_instr(vcpu
, kvm_vcpu_trap_il_is32bit(vcpu
));
1985 * The IPA is reported as [MAX:12], so we need to
1986 * complement it with the bottom 12 bits from the
1987 * faulting VA. This is always 12 bits, irrespective
1990 fault_ipa
|= kvm_vcpu_get_hfar(vcpu
) & ((1 << 12) - 1);
1991 ret
= io_mem_abort(vcpu
, run
, fault_ipa
);
1995 /* Userspace should not be able to register out-of-bounds IPAs */
1996 VM_BUG_ON(fault_ipa
>= kvm_phys_size(vcpu
->kvm
));
1998 if (fault_status
== FSC_ACCESS
) {
1999 handle_access_fault(vcpu
, fault_ipa
);
2004 ret
= user_mem_abort(vcpu
, fault_ipa
, memslot
, hva
, fault_status
);
2008 srcu_read_unlock(&vcpu
->kvm
->srcu
, idx
);
2012 static int handle_hva_to_gpa(struct kvm
*kvm
,
2013 unsigned long start
,
2015 int (*handler
)(struct kvm
*kvm
,
2016 gpa_t gpa
, u64 size
,
2020 struct kvm_memslots
*slots
;
2021 struct kvm_memory_slot
*memslot
;
2024 slots
= kvm_memslots(kvm
);
2026 /* we only care about the pages that the guest sees */
2027 kvm_for_each_memslot(memslot
, slots
) {
2028 unsigned long hva_start
, hva_end
;
2031 hva_start
= max(start
, memslot
->userspace_addr
);
2032 hva_end
= min(end
, memslot
->userspace_addr
+
2033 (memslot
->npages
<< PAGE_SHIFT
));
2034 if (hva_start
>= hva_end
)
2037 gpa
= hva_to_gfn_memslot(hva_start
, memslot
) << PAGE_SHIFT
;
2038 ret
|= handler(kvm
, gpa
, (u64
)(hva_end
- hva_start
), data
);
2044 static int kvm_unmap_hva_handler(struct kvm
*kvm
, gpa_t gpa
, u64 size
, void *data
)
2046 unmap_stage2_range(kvm
, gpa
, size
);
2050 int kvm_unmap_hva_range(struct kvm
*kvm
,
2051 unsigned long start
, unsigned long end
)
2056 trace_kvm_unmap_hva_range(start
, end
);
2057 handle_hva_to_gpa(kvm
, start
, end
, &kvm_unmap_hva_handler
, NULL
);
2061 static int kvm_set_spte_handler(struct kvm
*kvm
, gpa_t gpa
, u64 size
, void *data
)
2063 pte_t
*pte
= (pte_t
*)data
;
2065 WARN_ON(size
!= PAGE_SIZE
);
2067 * We can always call stage2_set_pte with KVM_S2PTE_FLAG_LOGGING_ACTIVE
2068 * flag clear because MMU notifiers will have unmapped a huge PMD before
2069 * calling ->change_pte() (which in turn calls kvm_set_spte_hva()) and
2070 * therefore stage2_set_pte() never needs to clear out a huge PMD
2071 * through this calling path.
2073 stage2_set_pte(kvm
, NULL
, gpa
, pte
, 0);
2078 int kvm_set_spte_hva(struct kvm
*kvm
, unsigned long hva
, pte_t pte
)
2080 unsigned long end
= hva
+ PAGE_SIZE
;
2081 kvm_pfn_t pfn
= pte_pfn(pte
);
2087 trace_kvm_set_spte_hva(hva
);
2090 * We've moved a page around, probably through CoW, so let's treat it
2091 * just like a translation fault and clean the cache to the PoC.
2093 clean_dcache_guest_page(pfn
, PAGE_SIZE
);
2094 stage2_pte
= kvm_pfn_pte(pfn
, PAGE_S2
);
2095 handle_hva_to_gpa(kvm
, hva
, end
, &kvm_set_spte_handler
, &stage2_pte
);
2100 static int kvm_age_hva_handler(struct kvm
*kvm
, gpa_t gpa
, u64 size
, void *data
)
2106 WARN_ON(size
!= PAGE_SIZE
&& size
!= PMD_SIZE
&& size
!= PUD_SIZE
);
2107 if (!stage2_get_leaf_entry(kvm
, gpa
, &pud
, &pmd
, &pte
))
2111 return stage2_pudp_test_and_clear_young(pud
);
2113 return stage2_pmdp_test_and_clear_young(pmd
);
2115 return stage2_ptep_test_and_clear_young(pte
);
2118 static int kvm_test_age_hva_handler(struct kvm
*kvm
, gpa_t gpa
, u64 size
, void *data
)
2124 WARN_ON(size
!= PAGE_SIZE
&& size
!= PMD_SIZE
&& size
!= PUD_SIZE
);
2125 if (!stage2_get_leaf_entry(kvm
, gpa
, &pud
, &pmd
, &pte
))
2129 return kvm_s2pud_young(*pud
);
2131 return pmd_young(*pmd
);
2133 return pte_young(*pte
);
2136 int kvm_age_hva(struct kvm
*kvm
, unsigned long start
, unsigned long end
)
2140 trace_kvm_age_hva(start
, end
);
2141 return handle_hva_to_gpa(kvm
, start
, end
, kvm_age_hva_handler
, NULL
);
2144 int kvm_test_age_hva(struct kvm
*kvm
, unsigned long hva
)
2148 trace_kvm_test_age_hva(hva
);
2149 return handle_hva_to_gpa(kvm
, hva
, hva
, kvm_test_age_hva_handler
, NULL
);
2152 void kvm_mmu_free_memory_caches(struct kvm_vcpu
*vcpu
)
2154 mmu_free_memory_cache(&vcpu
->arch
.mmu_page_cache
);
2157 phys_addr_t
kvm_mmu_get_httbr(void)
2159 if (__kvm_cpu_uses_extended_idmap())
2160 return virt_to_phys(merged_hyp_pgd
);
2162 return virt_to_phys(hyp_pgd
);
2165 phys_addr_t
kvm_get_idmap_vector(void)
2167 return hyp_idmap_vector
;
2170 static int kvm_map_idmap_text(pgd_t
*pgd
)
2174 /* Create the idmap in the boot page tables */
2175 err
= __create_hyp_mappings(pgd
, __kvm_idmap_ptrs_per_pgd(),
2176 hyp_idmap_start
, hyp_idmap_end
,
2177 __phys_to_pfn(hyp_idmap_start
),
2180 kvm_err("Failed to idmap %lx-%lx\n",
2181 hyp_idmap_start
, hyp_idmap_end
);
2186 int kvm_mmu_init(void)
2190 hyp_idmap_start
= kvm_virt_to_phys(__hyp_idmap_text_start
);
2191 hyp_idmap_start
= ALIGN_DOWN(hyp_idmap_start
, PAGE_SIZE
);
2192 hyp_idmap_end
= kvm_virt_to_phys(__hyp_idmap_text_end
);
2193 hyp_idmap_end
= ALIGN(hyp_idmap_end
, PAGE_SIZE
);
2194 hyp_idmap_vector
= kvm_virt_to_phys(__kvm_hyp_init
);
2197 * We rely on the linker script to ensure at build time that the HYP
2198 * init code does not cross a page boundary.
2200 BUG_ON((hyp_idmap_start
^ (hyp_idmap_end
- 1)) & PAGE_MASK
);
2202 kvm_debug("IDMAP page: %lx\n", hyp_idmap_start
);
2203 kvm_debug("HYP VA range: %lx:%lx\n",
2204 kern_hyp_va(PAGE_OFFSET
),
2205 kern_hyp_va((unsigned long)high_memory
- 1));
2207 if (hyp_idmap_start
>= kern_hyp_va(PAGE_OFFSET
) &&
2208 hyp_idmap_start
< kern_hyp_va((unsigned long)high_memory
- 1) &&
2209 hyp_idmap_start
!= (unsigned long)__hyp_idmap_text_start
) {
2211 * The idmap page is intersecting with the VA space,
2212 * it is not safe to continue further.
2214 kvm_err("IDMAP intersecting with HYP VA, unable to continue\n");
2219 hyp_pgd
= (pgd_t
*)__get_free_pages(GFP_KERNEL
| __GFP_ZERO
, hyp_pgd_order
);
2221 kvm_err("Hyp mode PGD not allocated\n");
2226 if (__kvm_cpu_uses_extended_idmap()) {
2227 boot_hyp_pgd
= (pgd_t
*)__get_free_pages(GFP_KERNEL
| __GFP_ZERO
,
2229 if (!boot_hyp_pgd
) {
2230 kvm_err("Hyp boot PGD not allocated\n");
2235 err
= kvm_map_idmap_text(boot_hyp_pgd
);
2239 merged_hyp_pgd
= (pgd_t
*)__get_free_page(GFP_KERNEL
| __GFP_ZERO
);
2240 if (!merged_hyp_pgd
) {
2241 kvm_err("Failed to allocate extra HYP pgd\n");
2244 __kvm_extend_hypmap(boot_hyp_pgd
, hyp_pgd
, merged_hyp_pgd
,
2247 err
= kvm_map_idmap_text(hyp_pgd
);
2252 io_map_base
= hyp_idmap_start
;
2259 void kvm_arch_commit_memory_region(struct kvm
*kvm
,
2260 const struct kvm_userspace_memory_region
*mem
,
2261 const struct kvm_memory_slot
*old
,
2262 const struct kvm_memory_slot
*new,
2263 enum kvm_mr_change change
)
2266 * At this point memslot has been committed and there is an
2267 * allocated dirty_bitmap[], dirty pages will be be tracked while the
2268 * memory slot is write protected.
2270 if (change
!= KVM_MR_DELETE
&& mem
->flags
& KVM_MEM_LOG_DIRTY_PAGES
)
2271 kvm_mmu_wp_memory_region(kvm
, mem
->slot
);
2274 int kvm_arch_prepare_memory_region(struct kvm
*kvm
,
2275 struct kvm_memory_slot
*memslot
,
2276 const struct kvm_userspace_memory_region
*mem
,
2277 enum kvm_mr_change change
)
2279 hva_t hva
= mem
->userspace_addr
;
2280 hva_t reg_end
= hva
+ mem
->memory_size
;
2281 bool writable
= !(mem
->flags
& KVM_MEM_READONLY
);
2284 if (change
!= KVM_MR_CREATE
&& change
!= KVM_MR_MOVE
&&
2285 change
!= KVM_MR_FLAGS_ONLY
)
2289 * Prevent userspace from creating a memory region outside of the IPA
2290 * space addressable by the KVM guest IPA space.
2292 if (memslot
->base_gfn
+ memslot
->npages
>=
2293 (kvm_phys_size(kvm
) >> PAGE_SHIFT
))
2296 down_read(¤t
->mm
->mmap_sem
);
2298 * A memory region could potentially cover multiple VMAs, and any holes
2299 * between them, so iterate over all of them to find out if we can map
2300 * any of them right now.
2302 * +--------------------------------------------+
2303 * +---------------+----------------+ +----------------+
2304 * | : VMA 1 | VMA 2 | | VMA 3 : |
2305 * +---------------+----------------+ +----------------+
2307 * +--------------------------------------------+
2310 struct vm_area_struct
*vma
= find_vma(current
->mm
, hva
);
2311 hva_t vm_start
, vm_end
;
2313 if (!vma
|| vma
->vm_start
>= reg_end
)
2317 * Mapping a read-only VMA is only allowed if the
2318 * memory region is configured as read-only.
2320 if (writable
&& !(vma
->vm_flags
& VM_WRITE
)) {
2326 * Take the intersection of this VMA with the memory region
2328 vm_start
= max(hva
, vma
->vm_start
);
2329 vm_end
= min(reg_end
, vma
->vm_end
);
2331 if (vma
->vm_flags
& VM_PFNMAP
) {
2332 gpa_t gpa
= mem
->guest_phys_addr
+
2333 (vm_start
- mem
->userspace_addr
);
2336 pa
= (phys_addr_t
)vma
->vm_pgoff
<< PAGE_SHIFT
;
2337 pa
+= vm_start
- vma
->vm_start
;
2339 /* IO region dirty page logging not allowed */
2340 if (memslot
->flags
& KVM_MEM_LOG_DIRTY_PAGES
) {
2345 ret
= kvm_phys_addr_ioremap(kvm
, gpa
, pa
,
2352 } while (hva
< reg_end
);
2354 if (change
== KVM_MR_FLAGS_ONLY
)
2357 spin_lock(&kvm
->mmu_lock
);
2359 unmap_stage2_range(kvm
, mem
->guest_phys_addr
, mem
->memory_size
);
2361 stage2_flush_memslot(kvm
, memslot
);
2362 spin_unlock(&kvm
->mmu_lock
);
2364 up_read(¤t
->mm
->mmap_sem
);
2368 void kvm_arch_free_memslot(struct kvm
*kvm
, struct kvm_memory_slot
*free
,
2369 struct kvm_memory_slot
*dont
)
2373 int kvm_arch_create_memslot(struct kvm
*kvm
, struct kvm_memory_slot
*slot
,
2374 unsigned long npages
)
2379 void kvm_arch_memslots_updated(struct kvm
*kvm
, u64 gen
)
2383 void kvm_arch_flush_shadow_all(struct kvm
*kvm
)
2385 kvm_free_stage2_pgd(kvm
);
2388 void kvm_arch_flush_shadow_memslot(struct kvm
*kvm
,
2389 struct kvm_memory_slot
*slot
)
2391 gpa_t gpa
= slot
->base_gfn
<< PAGE_SHIFT
;
2392 phys_addr_t size
= slot
->npages
<< PAGE_SHIFT
;
2394 spin_lock(&kvm
->mmu_lock
);
2395 unmap_stage2_range(kvm
, gpa
, size
);
2396 spin_unlock(&kvm
->mmu_lock
);
2400 * See note at ARMv7 ARM B1.14.4 (TL;DR: S/W ops are not easily virtualized).
2403 * - S/W ops are local to a CPU (not broadcast)
2404 * - We have line migration behind our back (speculation)
2405 * - System caches don't support S/W at all (damn!)
2407 * In the face of the above, the best we can do is to try and convert
2408 * S/W ops to VA ops. Because the guest is not allowed to infer the
2409 * S/W to PA mapping, it can only use S/W to nuke the whole cache,
2410 * which is a rather good thing for us.
2412 * Also, it is only used when turning caches on/off ("The expected
2413 * usage of the cache maintenance instructions that operate by set/way
2414 * is associated with the cache maintenance instructions associated
2415 * with the powerdown and powerup of caches, if this is required by
2416 * the implementation.").
2418 * We use the following policy:
2420 * - If we trap a S/W operation, we enable VM trapping to detect
2421 * caches being turned on/off, and do a full clean.
2423 * - We flush the caches on both caches being turned on and off.
2425 * - Once the caches are enabled, we stop trapping VM ops.
2427 void kvm_set_way_flush(struct kvm_vcpu
*vcpu
)
2429 unsigned long hcr
= *vcpu_hcr(vcpu
);
2432 * If this is the first time we do a S/W operation
2433 * (i.e. HCR_TVM not set) flush the whole memory, and set the
2436 * Otherwise, rely on the VM trapping to wait for the MMU +
2437 * Caches to be turned off. At that point, we'll be able to
2438 * clean the caches again.
2440 if (!(hcr
& HCR_TVM
)) {
2441 trace_kvm_set_way_flush(*vcpu_pc(vcpu
),
2442 vcpu_has_cache_enabled(vcpu
));
2443 stage2_flush_vm(vcpu
->kvm
);
2444 *vcpu_hcr(vcpu
) = hcr
| HCR_TVM
;
2448 void kvm_toggle_cache(struct kvm_vcpu
*vcpu
, bool was_enabled
)
2450 bool now_enabled
= vcpu_has_cache_enabled(vcpu
);
2453 * If switching the MMU+caches on, need to invalidate the caches.
2454 * If switching it off, need to clean the caches.
2455 * Clean + invalidate does the trick always.
2457 if (now_enabled
!= was_enabled
)
2458 stage2_flush_vm(vcpu
->kvm
);
2460 /* Caches are now on, stop trapping VM ops (until a S/W op) */
2462 *vcpu_hcr(vcpu
) &= ~HCR_TVM
;
2464 trace_kvm_toggle_cache(*vcpu_pc(vcpu
), was_enabled
, now_enabled
);